I retired from personal blogging in July 2008 but you can find me over at blog.xero.com
Follow @roddrury
Sitting on NZ0450 to Foo Camp with all the usual Wellington Software suspects I read a good article in the latest ‘The Channel’ magazine. It was an explanation of Blade Servers by Gary Elmes of IBM.
As a software guy I don’t understand hardware at all, but yesterday I had a look at the new Intergen hosting facility in Wellington. I actually sighted a little server I had there.
Gary’s article explained that Blades were more than just thin servers.
In any server there are two types of components. Those that provide the essence of the server’s functionality - CPU, RAM, IO Bus, etc. The others provide supporting functions - power supplies, fans, KVM connections and so on. Some of these later components are bulky, hot and take up a lot of space. And in a traditional server environment they are duplicated in each device; taking up real-estate, often running at low levels of utilisation and provising more to go wrong.
In the Blade infrastructure the Blades themselves provide the essential components while the supporting components are relegated to the chassis.
Well, that’s the first time Blades have been explained to me. Makes a lot of sense.
blog
From JohnR (thanks for the detailed response):
Conceptually, blades are great. However, in reality I have experienced
several (pretty big) limitations:
1) Any hardware cost savings only kick in at high utilisation ratios,
i.e. when your blade chassis’ are at least 75% full. So the idea that
you aren’t duplicating power supplies, fans, etc. doesn’t actually
result in cheaper hardware.
2) Blades themselves, mainly due to their form factor, are a lot less
flexible than traditional rack-mount servers: most blades don’t have
PCI slots, have only 4 memory slots (forcing you to buy more expensive
2 and 4 GB memory modules), don’t have hot-swap hard drives, no USB
ports (e.g. for dongles), and the list goes on… HP have addressed
some of these issues with their recent family (the c-class), which is
significantly better than older HP models and anything I’ve seen from
IBM so far.
3) Server rooms are not designed to cope with the significantly higher
server density provided by blades, not even remotely close. If you
decide to replace existing rack-mount servers with blades, you will
end up with racks that are 70% empty, because your existing server
room can’t provide enough power, cooling and even floor strength.
Crunching down the size of the server doesn’t do anything for the
power and cooling requirements of each individual server, so –again–
you don’t actually gain a lot.
4) Hardware reliability becomes more important. As you’ve now crammed
up to 16 blades into a blade enclosure (or chassis), every component
of that enclosure (the power supplies, the cooling fans, the
backplane, the management interface, the remote control interface, the
networking components) now becomes a potential point of failure. Of
course, there is a lot of redundancy built in, and most of the time
this works fine. However, I have experienced a situation with a faulty
chassis, where diagnosing the problem proves difficult, and so you end
up having to shut down over a dozen machines every time you have to
e.g. “upgrade the firmware on the backplane” to try to fix an
intermittent problem you’re having with the enclosure.
Even though you’ve reduced the physical size required for each
individual server, I don’t see that as a very compelling argument for
blades, especially given the above limitations (and there are many
others). Of course, if you need maximum computing power in the least
possible amount of room, then yes, blades are ideal. The perfect use
of blades is right here in Wellington, at Weta Digital: they need as
many CPUs as they can cram into a room. However, they require their
own substation, liquid-cooled radiators on the doors of the racks, and
any time an air conditioner fails, they have massive heat problems and
have to shut down significant portions of the farm while the problem
is fixed.
And after all this, there remains the issue of utilisation: you have
more CPUs per cubic metre than ever before (and they are ever more
powerful), but the CPUs are still sitting idle while chewing power and
generating heat. Which is why I think virtualisation is a much better
proposition: keep using the existing rack-mount servers, with the
maturity and flexibility they bring, just run more than one Operating
System on them. Virtualisation is getting to be pretty mature, and has
real advantages to offer, mainly in terms of utilisation, but also
manageability, platform stability (your Virtual Machines always see
the same virtual hardware, no matter what physical hardware they are
running on), and redundancy and high availability (you can move
running Virtual Machines between physical hosts on the fly, with zero
downtime).
http://www.vmware.com/products/vi/overview.html has a nice overview of
the key points offered by virtualisation. The main thing holding
virtualisation back is a cultural one: people are afraid of it, don’t
understand it and trust it enough to deploy it in a big way in
production environments. This is changing, and in the next three years
there will be major shift towards virtualisation, as CPUs become ever
more powerful and under-utilised.
I hope the above doesn’t come across as a rant; I’ve been working with
blades for a couple of years now, and have yet to be convinced of
their usefulness in the majority of general situations.